A municipal government just shipped a frontier-class language model. IplanRIO, the IT company that runs Rio de Janeiro's digital services, posted Rio 3.5 Open 397B to Hugging Face this week, fully open under an MIT license. Yes, a city hall.
It is a post-train of Alibaba's Qwen 3.5-397B-A17B, not a model built from scratch. The architecture is Mixture-of-Experts: 397 billion total parameters, around 17 billion active, with a 1 million token context window. The headline trick is SwiReasoning, a training-free inference method that flips between visible chain-of-thought and silent latent-space reasoning based on entropy signals, which the team says cuts token use while lifting accuracy.
The benchmark claims are the catch. On its own model card, Rio edges Qwen 3.7 Plus on four of five coding tests, including Terminal-Bench 2.1 at 70.8 versus 70.3, and posts a hefty 18-point jump over the base model there. These are first-party numbers. Nobody has independently audited them yet, and the model loses to GPT 5.5 across most categories.
SwiReasoning itself comes from a research paper by Shi et al., published last October. The weights run roughly 807 GB, so running this locally isn't happening for most people. No inference provider hosts it yet.
Bottom Line
A city government's IT company released a 397B-parameter MoE model under MIT, though its benchmark wins over Qwen 3.7 Plus are self-reported and unaudited.
Quick Facts
- 397B total parameters, ~17B active (MoE)
- Base model: Qwen 3.5-397B-A17B (Alibaba)
- 1,010,000 token context window
- MIT license, ~807 GB weights
- Terminal-Bench 2.1: 70.8 (company-reported)
